search

sanweiliti / RoHM

The official PyTorch code for RoHM: Robust Human Motion Reconstruction via Diffusion.
https://sanweiliti.github.io/ROHM/ROHM.html
Other
277 stars 11 forks source link

About motion_repr_clean used in test process #12

Open buptxyb666 opened 1 month ago

buptxyb666 commented 1 month ago

Thanks for your great work！

Why is the motion_repr_clean data involved in the test process, and how should it be handled during the video demo inference process without GT ?

Looking forward to your reply.

sanweiliti commented 1 month ago

Hi,

the GT motions are only used in the test/eval script for AMASS dataset for the following purposes: 1) serve as input (occluded) if we assume no noise in the input sequence; 2) calculate the evaluation metrics and provide visualization comparison between predictions and GT.

I suggest you go into the details of test/eval scripts, particularly for PROX/EgoBody to see how RoHM reconstructs given the input sequence.

buptxyb666 commented 1 month ago

Thanks, one more question, why combine val_output_traj output from TrajNet and motion_repr_clean to construct new_motion_repr and obtain traj_rec_full. Why not replace noised traj directly?